diff --git a/README.md b/README.md
index e5cbe35..be1f095 100644
--- a/README.md
+++ b/README.md
@@ -112,8 +112,8 @@ python trace.py
 The core SDKs and most integrations capture inputs and outputs by default.
 To disable input and output capture, set `captureInput: false` and `captureOutput: false` in TypeScript `init()`.
 For Python, use `capture_input=False` and `capture_output=False`.
-Some integrations work differently: MCP captures no tool arguments or results unless you pass
-`capturePayloads: true`, the AI SDK integration follows AI SDK's `recordInputs` and `recordOutputs`,
+Some integrations work differently: MCP also drops tool arguments and results when you pass
+`capturePayloads: false`, the AI SDK integration follows AI SDK's `recordInputs` and `recordOutputs`,
 Cursor reads `captureInput` and `captureOutput` from its own config file, and TanStack AI has no
 capture setting of its own and records chat messages, tool payloads, and error messages.
 Disabling output capture omits message content; response IDs and models, finish reasons, usage,
diff --git a/packages/mcp/README.md b/packages/mcp/README.md
index 6d2d7ec..d566490 100644
--- a/packages/mcp/README.md
+++ b/packages/mcp/README.md
@@ -64,7 +64,7 @@ Request spans also carry `mcp.client.name` and `mcp.client.version`, from the `c
 
 ## Payload capture
 
-Tool arguments and successful tool results are not captured by default. Enable them with `instrumentMcpTransport(transport, { capturePayloads: true })`. Core `@telemetry-dev/sdk` capture and masking settings still apply.
+`tools/call` spans record the tool arguments and the tool result, including a result with `isError: true`. Core `@telemetry-dev/sdk` settings still apply: `captureInput: false` drops the arguments, `captureOutput: false` drops the result, and `mask` runs on both. To leave tool payloads out while the core SDK captures other content, pass `instrumentMcpTransport(transport, { capturePayloads: false })`. Error codes and the JSON-RPC error message are recorded either way; `captureOutput: false` and `mask` do not apply to them.
 
 Trace context propagates by default. OpenTelemetry baggage remains local unless you enable it with `instrumentMcpTransport(transport, { propagateBaggage: true })`.
 
diff --git a/packages/mcp/src/index.ts b/packages/mcp/src/index.ts
index aefeaec..483ecbd 100644
--- a/packages/mcp/src/index.ts
+++ b/packages/mcp/src/index.ts
@@ -18,6 +18,11 @@ interface McpTransport {
 }
 
 export interface InstrumentMcpTransportOptions {
+  /**
+   * Record tool call arguments and results, failed results included, on `tools/call` spans.
+   * Defaults to true; the core SDK's captureInput, captureOutput, and mask still apply. Set false
+   * to keep tool payloads out even when the core SDK captures content.
+   */
   capturePayloads?: boolean;
   propagateBaggage?: boolean;
 }
@@ -354,7 +359,8 @@ function finishResponse(
       setFailure(pending, "tool_error");
     }
 
-    if (pending.capturePayloads && pending.method === "tools/call" && result?.isError !== true) {
+    // A failed result's content says what went wrong, so it is kept like a successful one.
+    if (pending.capturePayloads && pending.method === "tools/call") {
       pending.handle.update({ output: message.message.result });
     }
   }
@@ -424,7 +430,7 @@ export function instrumentMcpTransport<T extends McpTransport>(
   const incoming = new Map<JsonRpcId, PendingRequest>();
   const sendingResponses = new Set<PendingRequest>();
   const dispatching = new WeakSet<object>();
-  const capturePayloads = options.capturePayloads === true;
+  const capturePayloads = options.capturePayloads !== false;
   const propagateBaggage = options.propagateBaggage === true;
   let onmessage = target.onmessage;
 
diff --git a/packages/mcp/tests/mcp.test.ts b/packages/mcp/tests/mcp.test.ts
index 69d8d43..7e74910 100644
--- a/packages/mcp/tests/mcp.test.ts
+++ b/packages/mcp/tests/mcp.test.ts
@@ -35,10 +35,13 @@ interface JsonRecord {
   [key: string]: JsonValue | undefined;
 }
 
-function setup(): InMemorySpanExporter {
+function setup(
+  capture: { captureInput?: boolean; captureOutput?: boolean } = {},
+): InMemorySpanExporter {
   const spans = new InMemorySpanExporter();
   init(
     {
+      ...capture,
       apiKey: "td_live_test",
       serviceName: "mcp-tests",
       environment: "test",
@@ -69,11 +72,11 @@ afterEach(async () => {
   await shutdown();
 });
 
-test("instruments both v2 transport directions with propagation and opt-in payloads", async () => {
+test("instruments both v2 transport directions with propagation and payloads", async () => {
   const spans = setup();
   const [clientRaw, serverRaw] = InMemoryTransport.createLinkedPair();
-  const client = instrumentMcpTransport(clientRaw, { capturePayloads: true });
-  const server = instrumentMcpTransport(serverRaw, { capturePayloads: true });
+  const client = instrumentMcpTransport(clientRaw);
+  const server = instrumentMcpTransport(serverRaw);
   expect(instrumentMcpTransport(client)).toBe(clientRaw);
 
   client.onmessage = () => undefined;
@@ -294,6 +297,57 @@ test.each([
   expect(serverSpan.status.code).toBe(serverFault ? SpanStatusCode.ERROR : SpanStatusCode.UNSET);
 });
 
+test.each([
+  { name: "capturePayloads false", capture: {}, payloads: false, args: false, result: false },
+  {
+    name: "core captureInput false",
+    capture: { captureInput: false },
+    payloads: true,
+    args: false,
+    result: true,
+  },
+  {
+    name: "core captureOutput false",
+    capture: { captureOutput: false },
+    payloads: true,
+    args: true,
+    result: false,
+  },
+])("$name keeps tool payloads out", async ({ capture, payloads, args, result }) => {
+  const spans = setup(capture);
+
+  const [client, server] = InMemoryTransport.createLinkedPair().map((transport) =>
+    instrumentMcpTransport(transport, payloads ? {} : { capturePayloads: false }),
+  );
+
+  client.onmessage = () => undefined;
+  server.onmessage = () => {
+    void server.send({
+      jsonrpc: "2.0",
+      id: 5,
+      result: { content: [{ type: "text", text: "ok" }] },
+    });
+  };
+
+  await client.start();
+  await server.start();
+  await client.send({
+    jsonrpc: "2.0",
+    id: 5,
+    method: "tools/call",
+    params: { name: "weather", arguments: { city: "Paris" } },
+  });
+
+  for (const span of await exported(spans, 2)) {
+    expect(span.attributes["gen_ai.tool.call.arguments"]).toBe(
+      args ? '{"city":"Paris"}' : undefined,
+    );
+    expect(span.attributes["gen_ai.tool.call.result"]).toBe(
+      result ? '{"content":[{"type":"text","text":"ok"}]}' : undefined,
+    );
+  }
+});
+
 test("marks CallToolResult errors and closes pending requests", async () => {
   const spans = setup();
 
@@ -317,6 +371,10 @@ test("marks CallToolResult errors and closes pending requests", async () => {
   const toolSpans = await exported(spans, 2);
   expect(toolSpans.every((span) => span.attributes["error.type"] === "tool_error")).toBe(true);
   expect(toolSpans.every((span) => span.status.code === SpanStatusCode.ERROR)).toBe(true);
+  expect(toolSpans.map((span) => span.attributes["gen_ai.tool.call.result"])).toEqual([
+    '{"content":[{"type":"text","text":"failed"}],"isError":true}',
+    '{"content":[{"type":"text","text":"failed"}],"isError":true}',
+  ]);
 
   await shutdown();
   const closeSpans = setup();